⚡ electrical engineering

Active Learning for Optimal Experimental Design in Machine Learning-Based Building Energy System Identification

This paper systematically evaluates fourteen active learning techniques for optimal experimental design in building energy system identification, demonstrating that they significantly outperform passive random sampling by reducing model error by up to 54% across various scenarios and model architectures.

Original authors: Nam T. Nguyen, Truong X. Nghiem

Published 2026-06-25

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Nam T. Nguyen, Truong X. Nghiem

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to teach a robot how to control the heating and cooling system of a building. The goal is to make the robot smart enough to predict how the room temperature will change so it can save energy. But here's the catch: you can't just let the robot guess randomly. Every time you test a setting (like turning the fan speed up or down), you use energy and time. You want to get the most "learning" out of every single test.

This paper is like a guidebook for the robot's teacher. It asks: "What is the smartest way to choose our next test?"

The Problem: Throwing Darts Blindfolded

Usually, when scientists train these AI models, they use a method called Passive Learning. Imagine you are trying to learn the layout of a dark room by throwing darts at the wall. In Passive Learning, you just throw the darts completely at random. Sometimes you hit a spot that teaches you something new, but often you hit the same spot twice or throw a dart at a blank wall that tells you nothing. It's inefficient and wastes your "dart budget."

The Solution: Active Learning (The Smart Teacher)

The authors propose using Active Learning. Instead of throwing darts randomly, the teacher looks at what the robot already knows and asks: "Where are we most confused? Let's throw a dart right there!"

The paper compares 14 different "smart teacher" strategies to see which one helps the robot learn the building's temperature dynamics the fastest and most accurately. They tested these strategies on two types of "brains" for the robot:

The Deterministic Brain (Neural Network): A standard AI that gives one specific answer (e.g., "The temp will be 72°F").
The Stochastic Brain (Gaussian Process): An AI that gives an answer and a confidence level (e.g., "The temp will be 72°F, but I'm only 50% sure").

The Four Types of "Smart Teachers"

The researchers grouped the 14 strategies into four families, using some fun metaphors:

The Map Explorers (Data Space):
- Analogy: Imagine you are exploring a new city. These teachers say, "Let's go to the neighborhood we haven't visited yet." They don't care about the robot's current knowledge; they just want to make sure every part of the map gets a visit.
- Verdict: Good for starting out, but sometimes they visit boring places just because they haven't been there yet.
The Doubters (Uncertainty):
- Analogy: These teachers look at the robot and say, "You seem really unsure about this corner of the building. Let's test it right there." They focus on the spots where the robot is most confused.
- Verdict: Very effective, especially when the robot is a beginner.
The Structure Builders (Information Gain):
- Analogy: These teachers care about the "rules" of the building. They say, "Let's test a setting that will help us figure out the fundamental laws of how this building heats up, rather than just guessing a temperature."
- Verdict: Great for understanding the "why" behind the data, especially when you have a bit more data to start with.
The Change Agents (Model Change):
- Analogy: These teachers ask, "If we test this specific setting, how much will it force the robot to rewrite its notebook?" They look for the test that causes the biggest "Aha!" moment.
- Verdict: Very powerful for quickly updating the robot's brain.

The Big Experiment

The team didn't just talk about this; they ran a massive simulation using a high-fidelity building simulator (called BOPTEST). They set up three different scenarios:

Scenario A: The robot starts with almost no data (2 samples) and has very strict rules on how fast it can change the temperature (tight constraints).
Scenario B: Same as A, but the rules are a little looser.
Scenario C: The robot starts with more data (10 samples) and has very loose rules.

The Results: What Worked Best?

The paper found that Active Learning generally beats Random Learning, but the "best" strategy depends on the situation:

When the robot is a total beginner (Scenario A): The "Smart Teachers" shined. They reduced the robot's mistakes (errors) by up to 54% compared to the random throwers.
- For the Stochastic Brain (the one that knows its confidence), the "Structure Builder" (Information Gain) and "Map Explorer" strategies were the winners.
- For the Deterministic Brain (the standard AI), the "Change Agents" (Model Change) and "Doubters" (Uncertainty) worked best.
When the robot already knows a bit (Scenario C): The advantage of being "smart" started to fade. Once the robot had a decent amount of data to start with, the random throwers (Passive Learning) caught up. In some cases, being too picky about which test to run next didn't help much more than just trying things out.

The Bottom Line

This paper tells us that if you are trying to train an AI to control a building's energy, don't just guess randomly. If you are starting with very little data, use a "Smart Teacher" strategy to pick your tests. It will save you time and energy and give you a much smarter robot.

However, the paper also warns that there isn't one single "magic bullet" strategy. The best approach changes depending on how much data you have and how strict your building's rules are. The most dramatic improvements happen when you are in the "data-scarce" zone—when you have very little information to start with.

Technical Summary: Active Learning for Optimal Experimental Design in Machine Learning-Based Building Energy System Identification

Problem Statement
Machine learning (ML) techniques are increasingly adopted to identify the dynamics of building energy systems (BESs), particularly HVAC thermal dynamics, due to their flexibility compared to physics-based modeling. However, the performance of these data-driven models is heavily dependent on the quality of training data. Standard approaches often rely on passive learning (PL) with uniformly random control inputs (e.g., PRBS, white noise), which can be inefficient, requiring massive datasets to accurately estimate nonlinear system dynamics. The paper addresses the need for Optimal Experimental Design (OED) to determine which experiments to conduct to collect the most informative data. While OED and Active Learning (AL) share underlying mechanisms, existing building energy studies have largely applied single, ad-hoc acquisition functions without a systematic comparison across different AL paradigms or ML model classes. Furthermore, many prior works focus on static prediction or classification rather than online dynamical system identification under realistic control constraints.

Methodology
The study proposes a systematic comparison of fourteen AL techniques applied to the identification of building thermal dynamics. The methodology is structured around the following components:

System Formulation: The problem is modeled as a discrete-time dynamical system where the goal is to learn an approximation of the unknown nonlinear function $f$ governing the transition from state $x_k$ and control input $u_k$ to the successor state $x_{k+1}$ .
Model Classes: Two distinct ML architectures are evaluated:
- Deterministic: A Feedforward Neural Network (NN) approximating the dynamics via point predictions.
- Stochastic: A Gaussian Process (GP) providing a posterior predictive distribution with native uncertainty estimates.
Active Learning Taxonomy: The fourteen techniques are categorized into four groups based on their acquisition function logic:
- Data Space: Methods relying on geometric structure (Greedy Sampling, Inverse Distance Weighting).
- Uncertainty: Methods targeting regions of high predictive uncertainty (Query-by-Committee, Monte Carlo Dropout for NN; Maximize Variance, Integrated Variance Reduction for GP).
- Information Gain: Methods targeting hyperparameter estimation (Fisher Information, Shannon Entropy for GP).
- Model Change: Methods targeting the largest expected update to model parameters (Expected Model Change Maximization, Maximize Last Layer Change for NN; Upper Confidence Bound, Probability of Improvement for GP).
Experimental Setup: All algorithms are implemented and evaluated on the BOPTEST high-fidelity simulator, which uses Modelica to model realistic HVAC dynamics. The study employs an online learning framework with batch updates (retraining every 10 samples).
Test Scenarios: Performance is assessed across six scenarios varying in:
- Initial Dataset Size: 2 data points (data-scarce) vs. 10 data points.
- Control Constraints: Tight ramp constraints (simulating realistic HVAC limits) vs. wide ramp constraints.
- Evaluation Metrics: Root Mean Square Error (RMSE) computed over three time windows (0–2h, 0–12h, 0–24h) and reported as both windowed mean and final time-step values.

Key Contributions
The paper makes three primary contributions:

Systematic Taxonomy: It formally classifies and derives acquisition functions for fourteen AL techniques across both deterministic (NN) and stochastic (GP) model classes, organized into the four categories mentioned above.
High-Fidelity Evaluation: Unlike prior studies relying on static datasets or simple virtual testbeds, this work implements and evaluates all techniques on the BOPTEST simulator, incorporating realistic HVAC dynamics and control constraints.
Comprehensive Empirical Comparison: The study provides a broad empirical analysis across varying initial data sizes and operational constraints, benchmarking all AL methods against a passive learning baseline with uniformly random inputs.

Results
The simulation results indicate that AL-based models generally outperform passive learning (PL) with uniformly random control inputs, achieving error reductions of up to 54% in specific scenarios. However, the magnitude and consistency of these improvements vary significantly by acquisition function, model class, and operating regime:

Data-Scarce Regime (2 initial points, tight constraints): AL methods show the most significant benefits.
- For GP models, Integrated Variance Reduction (IVR) and Inverse Distance Weighting (IDWGP) performed strongly, with IDWGP achieving a 54% RMSE reduction in the 0–12h window.
- For NN models, Query-by-Committee (QBC) and gradient-based Model Change methods (EMCM, MLLC) provided the largest improvements, with QBC reducing mean RMSE by 27–42% in early windows.
Data-Rich Regime (10 initial points, loose constraints): The advantage of AL over PL diminishes. In several medium-to-long horizon scenarios, PL performed comparably to or better than specific AL methods (e.g., GSGP and SE for GP; MCD for NN).
Model-Specific Observations:
- Gaussian Processes: The Information Gain category (Fisher Information) became the strongest performer once ramp constraints were moderately loosened. The Data Space method GSGP consistently failed to improve over PL in all scenarios.
- Neural Networks: Uncertainty Sampling (specifically QBC) remained the most consistently competitive category. Monte Carlo Dropout (MCD) failed to improve over PL in the data-scarce scenario, suggesting calibration issues in this context.
General Trend: The benefits of OED are most pronounced and reliable when the initial training data is scarce and the admissible control range is tightly constrained.

Significance and Claims
The paper claims that Active Learning-based experimental design is a valuable tool for building energy system identification, particularly in the data-scarce regime where collecting data is expensive or time-consuming. It establishes that no single acquisition function is universally superior; the optimal choice depends on the model class (deterministic vs. stochastic) and the specific operational constraints (data availability and control ramp limits).

The study serves as a baseline benchmark for future research, highlighting that while AL can significantly reduce model error (up to 54% in the best cases), its benefits are not guaranteed in all conditions. The authors modestly conclude that their work focuses on a relatively simple system (four inputs, one output) and does not assess computational costs, suggesting that future work should extend these findings to more complex nonlinear systems and include statistical analysis of computational overhead.